Re: [BUG] net_cls: Panic occured when net_cls subsystem use

On Mon, Jun 01, 2009 at 06:06:03PM -0400, jamal wrote:

On Mon, 2009-06-01 at 22:49 +0200, Jarek Poplawski wrote:

quoted

Actually, I'd insist with the old rock and handling that other rude
u32 case, at least until it's fixed in place. So I attach my version
of your patch (additionally I removed a pair of braces because of
checkpatch warning).

Sure, it doesnt complicate anything - Minoru, this is the version to go
for.

quoted

Alas, I still think we don't need to change so much in -stable to
fix the cls_cgroup oops, so I attach a patch which I think is
enough for -stable and probably -net too. It could be "reverted"
in -net-next just after applying cls_api patch. Of course, treat
it only as my humble proposal, and feel free to recommend to David
your version, no problem (really).

My view is the same - that the second patch doesnt fix the root
cause; and it is not that complicated to fix the root cause. So I humbly
disagree with you.

The issue is that a classifer (cls_group in this example) is being
misconfigured. It rejects the config - but the tp has already been
added. 
It then tries to use the tp in the fast and fails. 

If you look as closely as you did with the patch i posted, youd find
ways to construct similar hostile misconfigs for other classifiers.
You just need to create the scenario where the attributes will fail to
validate. 

I actually suspect the most common scenario for such a failure is not
that head is null (I doubt in Minoru case that allocation will fail);
rather it is some reference to head->something.

The patch #2 is obviously worse and fixes less (of course it still
needs testing for Minoru's case), but I'm 100% confident it can't
introduce any regression (neither take 1 nor 2), which is much harder
to say about patch #1, considering various "rude" configs we could
miss (but we could give them some time to show off). But, as I've
written before, I'm really (really) OK with your decision.

Cheers,
Jarek P.